Ant releases and open-sources trillion-parameter reasoning model Ring-1T, approaching GPT-5 capabilities
Ring-1T: Ant Group’s Trillion-Parameter Open-Source Reasoning Model
In the early hours of October 14, Ant Group officially released Ring-1T, a trillion-parameter reasoning model with fully open-sourced model weights and training recipes.
This release builds upon the Ring-1T-preview version from September 30, extending large-scale Verifiable Reward Reinforcement Learning (RLVR) to boost natural language reasoning capabilities. The team also refined general model performance through RLHF training, achieving balanced results across multiple benchmarks.
---
Breaking Ground in Mathematical Reasoning
To advance Ring-1T’s mathematical and complex reasoning skills, the Bailing Team tested it on difficult IMO 2025 (International Mathematical Olympiad) problems:
- Framework: Integrated into AWorld, a multi-agent reasoning framework.
- Approach: Solutions generated purely via natural language reasoning.
- Results:
- Solved Problems 1, 3, 4, and 5 in a single attempt → IMO Silver Medal equivalent
- Third attempt achieved nearly full-score geometric proof reasoning for Problem 2
- On challenging Problem 6, matched Gemini 2.5 Pro’s answer (4048) — correct answer was 2112
---
General Capability Benchmarks
Ring-1T also demonstrated strong general abilities:
- Arena-Hard V2:
- Human Preference Alignment score: 81.59%
- #1 open-source ranking, close to GPT-5-Thinking (High) at 82.91%
- HealthBench:
- Top score among open-source healthcare QA systems
Performance comparison: Ring-1T vs. other reasoning models
---
Overcoming the Precision Divergence Challenge
Training-Inference Precision Mismatch
One of the largest hurdles in trillion-parameter model training is precision divergence — small implementation differences between training and inference that cause accuracy drops or even training collapse.
Ant Group’s “Icepop” Algorithm
Icepop uses masked bidirectional truncation to “freeze” distribution differences at a low level, ensuring stable long-sequence, long-duration training.
---
Scaling Reinforcement Learning with ASystem & AReaL
For large-scale RL on trillion-parameter models, Ant Group created ASystem, a high-performance RL platform with the open-source AReaL framework. Key optimizations include:
- Instant GPU memory fragment recovery
- Zero-redundancy weight swapping
- Billion-scale RL training stability — sustainable daily operations
(Left: GRPO divergence grows exponentially; Icepop remains stable. Right: GRPO divergence peaks sharply; Icepop stays low.)
---
Leveraging the Ling 2.0 Architecture
Ring-1T builds on the Ling 2.0 trillion-parameter base model, featuring:
- Highly sparse MoE architecture (1/32 expert activation)
- FP8 mixed precision
- MTP and other efficiency features
Post-training process:
- LongCoT-SFT
- RLVR
- RLHF
This process enhanced complex reasoning, instruction-following, and creative writing.
Users can download the model via HuggingFace or ModelScope, and test it online through Ant’s Baibaoxiang platform.
---
Expanding the Trillion-Parameter Model Family
Ant Bailing’s program now includes 18 models, from 16B parameters to 1T parameters, with two trillion-scale models:
- Ling-1T — General language model
- Ring-1T — Reasoning-focused model
This marks the program’s official entry into the 2.0 phase.
---
Beyond Reasoning: Monetizing AI Content
In the growing open-source AI ecosystem, Ring-1T demonstrates how openness can drive cutting-edge reasoning.
For creators aiming to monetize AI-generated content globally, AiToEarn官网 provides a fully open-source AI content monetization framework:
- Cross-platform publishing: Douyin, Kwai, WeChat, YouTube, Instagram, X (Twitter), and more
- Integrated tooling: AI generation, publishing, analytics
- Ranking visibility: AI模型排名
This enables creators to efficiently turn AI-powered creativity into sustainable income.
---
Would you like me to create a compact infographic-style summary of Ring-1T’s capabilities and unique features within this same Markdown file so it works as a quick reference? That could make the document even more reader-friendly.
